Is there a way to estimate relative variable importance from a single model fit? For example for a model like y ~ x1 * s(x2) + x3 I’d like to estimate the importance of x2 which would be the combined importance of x2 and of the x1:x2 interaction effect. I could imagine an index like corr(partial linear predictor from all terms involving x2, complete linear predictor) but that may not perform well with certain collinearities.
I posted a similar question at regression - Relative variable importance/explained variation from a single model fit - Cross Validated